Artificial Intelligence for Dietary Intake Measurement: A Scoping Review of Accuracy, Equity Gaps, and Implications for Underrepresented Populations

Background and Purpose: Accurate dietary intake measurement is foundational to identifying nutrition-related need, targeting interventions, and informing food assistance policy. However, conventional methods including 24-hour dietary recalls, food frequency questionnaires, and dietary records depend heavily on self-report capacity, placing a disproportionate burden on older adults with cognitive or physical limitations, young children, individuals with disabilities, and low-income or low-literacy groups. This systematic underrepresentation contributes to misclassification of nutritional need and inequitable allocation of nutrition services. AI-based approaches utilizing food image recognition, wearable sensors, and multimodal data inputs offer a lower-burden alternative, yet whether these tools perform accurately and equitably across underrepresented populations remains unclear. This scoping review synthesizes the current evidence on AI applications for dietary intake measurement and examines implications for equity-focused social and behavioral health practice.

Methods: Following PRISMA-ScR guidelines, we searched PubMed, Web of Science, Cochrane Library, and EBSCO from database inception through February 10, 2026. Eligible studies included experimental and observational designs applying machine learning, deep learning, or reinforcement learning to measure food or nutrient intake in human populations. Studies using non-AI rule-based methods, non-dietary inputs, or non-intake outcomes were excluded. Data were extracted on study design, sample characteristics, input modalities, analytic approaches, measurement performance outcomes, and equity considerations.

Results: 105 studies published between 2010 and 2026 met inclusion criteria, with sample sizes ranging from 10 to 38,415 participants. AI tools spanned diverse input modalities, including food and plate image recognition, wearable sensors capturing chewing and jaw motion, and text-based inputs, with some studies integrating multiple data streams. Food detection accuracy ranged from 74% to 99.85% across studies. For instance, an RGB-D fusion network achieved a mean absolute error of 15% in calorie estimation, and a sound-based model detected food intake through chewing patterns with up to 94% accuracy. Passive and sensor-based modalities, which require minimal active participation, suggest particular promise for populations unable to self-record. However, a critical equity gap emerged: most studies were conducted in controlled settings with homogeneous samples, and very few validated AI performance among older adults, children, individuals with disabilities, or racial and ethnic minority groups whose dietary patterns and food environments diverge substantially from those of majority study populations.

Conclusions and Implications: AI-based dietary intake measurement demonstrates strong technical performance; however, its equity potential remains largely unvalidated. Accurate, low-burden measurement tools could strengthen identification of nutrition-related need among marginalized populations currently undercounted by conventional methods, forming a critical foundation for future nutrition assessment, planning, food assistance targeting, and social welfare policy. Realizing this potential requires prioritizing algorithmic fairness, cultural dietary validity, and digital access equity in future research. Social work researchers are well positioned to lead community-engaged validation studies that center the populations most at risk of being further marginalized by poorly calibrated AI systems.